History Problem Core Idea Survey Results Impact Quiz Takeaways
Interactive Paper Explainer

A Map for the Agent Rush
CoALA

By 2023, LLM agents were proliferating without a vocabulary. CoALA borrowed seventy years of cognitive science to give them one: a decision cycle, four memory types, and a structured action space — then organized 100+ papers onto the map.

Start Learning Read the Paper ↗
4
Memory types
2
Action spaces
100+
Papers organized
2023
Sumers et al.
History

Two Traditions Collide

Symbolic AI built agents for forty years; LLMs rebuilt them in two — CoALA is the bridge.

1970s-90s
Symbolic cognitive architectures
SOAR, ACT-R, and BDI agents: explicit memories, deliberation cycles, carefully engineered modules — capable in narrow worlds, brittle everywhere else.
2017-22
LLMs eat everything
A single pretrained network replaces stacks of modules — but agent designs become folklore: prompts, scripts, and demos with no shared vocabulary.
2023
The agent explosion
ReAct, Reflexion, Toolformer, AutoGen… dozens of systems, each reinventing memory, control, and actions under its own names (entries #53, #55, #54, #107).
Sep 2023
🚀 CoALA
Sumers et al. map the territory with cognitive science's tools: decision cycle (retrieve/decide/act), modular memory, internal + external action spaces — retrospective survey + prospective research agenda.
2024+
The vocabulary sticks
Papers now self-locate on the CoALA map; the memory modules it named power the X-category systems (entries #97-104) and the agent frameworks it organized (XI).
The Two-Frame Synthesis

CoALA describes every language agent with two inherited frames. Frame 1 — the decision cycle: agents loop perceive context → retrieve relevant memory → reason/decide → act, where "act" splits into internal actions (memory reads/writes, reasoning steps) and external actions (tools, environments, communication). Frame 2 — the memory taxonomy: working (the immediate context), episodic (specific past experiences), semantic (facts/knowledge — including retrieved corpora), and procedural (skills and policies — including the model's own weights and code libraries). Any agent, paper or product, becomes a configuration within this design space.

Chapter 01

A Field with No Map

The 2023 problem: agents everywhere, structure nowhere.

🌀
Design as Folklore
  • Every new agent system reinvents its own terms: memory here is 'a scratchpad', there 'a vector DB', elsewhere unnamed
  • Comparisons are anecdotal — mechanism-level discussion impossible without shared parts
  • Progress is invisible: a new system's real novelty (a memory op? an action type?) is buried in prompts
  • Seventy years of cognitive-science and symbolic-AI design knowledge go unused
🗺
The CoALA Answer
  • A single architecture vocabulary: decision cycle + memory modules + action-space structure
  • Retrospective: 100+ recent agent papers organized onto the map — what exists, what's missing
  • Prospective: actionable gaps identified (procedural learning, memory management, self-directed growth)
  • Cognitive science as design library rather than nostalgia
Analogy — The Periodic Table Moment

Pre-CoALA agent design was alchemy: each lab had its own apparatus and secret terms, and results didn't transfer. CoALA is the periodic table — not a new element, but the grid that makes every element (including future ones) locatable, comparable, and combinable. Suddenly 'agent X adds episodic writes to a ReAct loop' is a sentence with exact meaning.

Chapter 02

The Decision Cycle

The loop every CoALA agent runs — with memory and action spaces as first-class parts.

1️⃣ Perceive context
Input, environment observations, and the current working memory define the agent's situation.
2️⃣ Retrieve
Relevant episodic and semantic memory is selected — retrieval into working memory is itself an internal action.
3️⃣ Reason & decide
The agent deliberates over retrieved content and procedural knowledge, selecting the next action (or reasoning step).
4️⃣ Act — internal or external
External: tools, environment, communication. Internal: memory writes, retrieval, self-directed reasoning. The cycle repeats.
The four memories
  • Working: the active context — the prompt itself, bounded and volatile
  • Episodic: specific past experiences (trials, conversations) — Reflexion's lesson buffer (entry #55)
  • Semantic: facts and knowledge — parametric weights, vector stores, corpora (RAG's domain, entry #29)
  • Procedural: skills and policies — the model itself, code, skills libraries (Voyager's)
The action space
  • External actions: affecting the world — tool calls, environment steps, messages to other agents
  • Internal actions: affecting the agent itself — memory writes, retrieval queries, reasoning chains, (re)planning
  • The split replaces vague "the agent thought about it" with an exact operation count
  • Agent maturity ≈ how much of the loop is internal (self-directed) vs scripted
Interactive Demo — One Agent, Two Vocabularies

The same ReAct-style agent described twice — folklore version vs CoALA version. Tab through the transformation.

Chapter 03

What the Survey Found

Locating the 2023 ecosystem on the map — the strengths and the empty quadrants.

The Retrospective Lens

The prospective half turned each gap into a research program — and the X-category memory papers (entries #97-104) read like its checklist executed.

Interactive Demo — The Decision Cycle, Annotated

Follow one agent turn through the CoALA loop — every stage named, every memory touched, both action kinds visible.

Chapter 05

A Vocabulary as a Result

CoALA's output is a map, not a metric — its impact is legibility.

FRAMEWORK
1 architecture
decision cycle + memory + action space
PAPERS MAPPED
100+
the 2023 agent ecosystem, organized
GAPS NAMED
procedural +
memory management, self-directed growth
ADOPTION
vocabulary
agent papers now self-locate on the map
Interactive Demo — The Empty Quadrant

CoALA's most-cited finding: one memory type lags the other three badly. Press reveal.

Memory typeContentLLM-agent implementation (2023 examples)
Workingactive contextthe prompt / scratchpad (ReAct traces, entry #53)
Episodicspecific experiencesReflexion's reflection buffer (entry #55)
Semanticfacts / knowledgeRAG corpora + vector stores; parametric weights (entry #29)
Proceduralskills / policiesmodel weights, skill libraries (Voyager) — the weak quadrant

The taxonomy with the paper's own example mappings — the vocabulary the whole agent-memory category now uses.

Legacy

Legacy — The Lingua Franca

CoALA became how the field describes agents — quietly, permanently.

🗣 The shared vocabulary
Memory-type and action-space terms from CoALA now appear as the standard descriptive layer in agent papers — the survey that became a dictionary.
📋 The research checklist
Its named gaps (procedural learning, episodic consolidation, self-directed growth) became the agenda the memory category executed (entries #97-104).
🏛 Cognitive science rehabilitated
The paper proved the old architectures' vocabulary was the missing layer — not their mechanisms — resetting the relationship between the two AI traditions.
🎓 Teaching infrastructure
CoALA is the framework through which agents are now taught — the mental model that makes 'agent' a structured concept rather than a demo genre.
⚠️ What it did NOT solve
It is descriptive, not normative — no benchmarks, no training recipe; the maps' borders (e.g. semantic vs procedural in a learned model) stay contested; and LLM agents keep producing structures the frame must stretch to fit.
🛤 Read next
The map's territories: ReAct · Reflexion · Memory Mechanism Survey
Test Yourself

Quick Quiz

Check your understanding of the key concepts from CoALA.

Reference

Key Takeaways

Everything you need to remember about this paper.

✅ CoALA = decision cycle + modular memory + structured action space: one map for all language agents.
✅ Four memories: working (prompt), episodic (experiences), semantic (facts/RAG), procedural (skills/weights).
✅ Internal actions (memory ops, reasoning) vs external actions (tools, environment) — self-management becomes designable.
✅ 100+ papers retrospectively organized; gaps prospectively named (procedural learning above all).
✅ The taxonomy the agent-memory category (X) uses as its native language.
✅ Read it as the field's periodic table — not a new element, the grid.